What Linear Probes Miss: Multi-View Probing for Weight-Space Learning

The explosive growth of open-source model repositories has created a Model Jungle, where checkpoints are frequently shared without adequate documentation or metadata. While weight-space learning offers a pathway to identify and analyze these models directly from their parameters, processing full-scale weights is computationally prohibitive. Probing-based methods have emerged as a lightweight alternative, extracting permutation-equivariant representations via learnable probe vectors. However, existing probing methods are limited by a single-view design: they capture first-order structures but fail to encode the rich, higher-order correlation patterns inherent in row--column interactions. To bridge this gap, we introduce MVProbe, a multi-perspective probing framework that synthesizes first-order signals with interaction-aware (Gram-based) views. Our approach is theoretically grounded; we analyze the scaling laws of different probing orders to derive a principled standardization and fusion strategy that ensures balanced contributions from all branches. On the Model Jungle benchmark, MVProbe consistently outperforms the state-of-the-art ProbeX across diverse architectures, including ResNet, SupViT, MAE, and DINO.

Each AI model stores its learned knowledge as a large matrix called weights. Is analyzing these matrices from a single view enough to identify what a model does? We discovered that the answer is no: single view creates a fundamental blind spot where two entirely different models can appear identical. MVProbe resolves this by examining weights from four complementary perspectives at once, capturing structure that a single view inherently misses. Every day, millions of AI models are shared online — but nearly one in four comes with no description, making it difficult to know what a model does without running it yourself. Our results show that MVProbe bring us closer to a world where finding the right model is as simple as a search.